Should Jev replace my extractor LLM?

No. Jev fits decision steps like classify, verify, dedupe, and route, while open-ended generation stays with a full LLM and certain cases stay in deterministic code.

What does the one built demo actually prove?

That swapping only the judge layer gives every same-or-different decision a calibrated confidence, with uncertain pairs going to a review queue instead of merging blindly.

Are the other backlog use cases validated?

No. The backlog states plainly that candidates like triple grounding, Cypher safety gates, and reranking are unvalidated until someone builds the small demo, except the marked dedup one.

What is the catch with calibration?

It is context-sensitive: state contents and question phrasing are part of the measurement, so batch related questions per call, scope each question explicitly, and tune thresholds against the batching you ship.

Which demo in the repository is actually built?

One: entity resolution and dedup under demos/dedup_jev. It ports neo4j-field/dedup-demo with the same three-phase pipeline and dirty dataset, changing only the judge layer to Jev.

How should I choose between a rule, Jev, and a full LLM?

Use a deterministic rule when the case is certain, Jev when it is a judgment call with a confidence gate, and escalate to a full LLM only what Jev is unsure about.

What does the extraction cascade use case propose?

A cheap model extracts, Jev scores each field or triple for probability of error, and only low-confidence records go to an expensive reasoning model. The README cites a public benchmark where routing Jev's ~10% uncertain cases matched frontier quality at ~17% of the cost.

CommunityJev Tools & CLIs

Jev Graphrag

Self-contained Neo4j demos showing where Jev helps GraphRAG pipelines: calibrated Choice, Score, and yes-no judgments on the decision steps, with open-ended generation left to a full LLM.

See highlights

Open github.com

Stars: 1
Language: Python
Published by: Community
Author: neo4j-field
Last verified: Oct 2026

neo4j-field/jev-graphrag

Organization repository on GitHub

View on GitHub

Small demos + use-case backlog: TypeSafe AI's Jev as a calibrated decision layer for GraphRAG pipelines on Neo4j.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Small, self-contained demos exploring how TypeSafe AI's Jev model (a "System One" model — calibrated Choice/Score/yes-no probability answers over a state, no free-text generation) can improve knowledge-graph extraction and GraphRAG pipelines on Neo4j.

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Rule of thumb throughout: deterministic rules for certain cases, Jev with a confidence gate for judgment calls, LLM escalation for the unsure.
Built demo ports a dirty-dataset entity dedup pipeline, keeping the same merges while routing low-confidence pairs to review.
Each demo runs independently after syncing deps and setting the API key plus Neo4j credentials; a narrated notebook walks every decision.
Documents how question scoping and state design affect calibration, and warns that confidential data needs retention review.

Watch out

Field-team demos; only the dedup pipeline is built and the backlog is unvalidated. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-deep-dive

Long-form Jev writeup that maps each conclusion to official docs, vendor evals, integrations, community reproductions, and press, separating contract facts from unverified claims.